Back-propagation neural network model enhanced by transfer learning for the prediction method of compressive strength of alkali residue-based foamed concrete

Abstract

This study introduces a novel artificial intelligence method for predicting the compressive strength (CS) of alkali residue-based foamed concrete (A-FC). The goal was to accurately predict the CS of A-FC in data-limited scenarios using key parameters such as density, water-to-cement ratio (W/C), and sand-to-cement ratio (S/C). This research employed a back-propagation neural network initially trained on CS data from 190 conventional foamed concrete (C-FC) groups (source domain data: W/C from 0.26 to 0.85, S/C from 0 to 4.29, and density from 0.43 to 2.00 g/cm³). The model then underwent transfer learning (TL) to improve its predictive capabilities. After adapting to 30 A-FC datasets (target domain data: W/C from 0.75 to 2.00, S/C from 0.25 to 2.33, and density from 0.35 to 0.90 g/cm³), the model demonstrated remarkable accuracy with new data. Its robustness was further validated by predicting the CS of an additional 12 A-FC groups. The results show a strong alignment between the model’s predicted values and the experimental data, highlighting its excellent predictive accuracy and generalization capabilities. Compared to traditional experimental methods, this approach significantly reduces both time and costs. The study examines the factors influencing the CS of C-FC and A-FC, improving the accuracy and efficiency of A-FC CS predictions and providing theoretical support for its large-scale application.